What is Artificial Intelligence?
Is a master of intelligent dialogue scene. Language is a very important aspect of embodied intelligence, language about equal intelligence, Turing test is based on language. In life, many of our mobile phone app has been applied artificial intelligence techniques such as: face recognition (graphics), speech recognition, personalized recommendations, intelligent scheduling, machine translation artificial intelligence technology has deeply affected our lives, we have no artificial intelligence can not imagine what the world is.
There are two questions that we have to think about:
The current development of artificial intelligence to what extent? To develop and to what extent?
The current weak AI Artificial Intelligence technology is still in the stage, we want to develop to the point? It is generally considered artificial intelligence development goal is to create people with the same agent, but to better answer this question, we must first understand what the answer to this question below Yes.
What is intelligence?
As an example, a company have not yet learned to walk with a baby has many powerful features of robots, we tend to think that the baby is smart, but the robot is not, why?
This concept is very vague intelligence, there are a lot of controversy, from the following two points, three levels to explain what is smart
Two points
Adaptive capacity: the current AI in a certain extent, knowledge of migration; a very important goal is to achieve universal AI Artificial Intelligence self-consciousness: "I know what they are doing", how to determine whether a self-conscious thing? (This is also why many people think that having a baby does not have the intelligence and robotics, currently has a kind of testing is to look in the mirror)
Three levels
1. Intelligent operation: Math
2. Intelligent Perception: visual, auditory
It has been basically solved, speech recognition and computer vision technology is already very mature
3. Cognitive Intelligence: logical reasoning, knowledge, understanding, thinking decisions
Language is a typical, very important application, language is a world of a system of symbols, either natural language or mathematical language, logic, language, programming language, language is the core of cognitive intelligence, if we can not express this world, we will not be able to change it
Other important applications as well as face recognition, posture recognition
Turing Test
Turing test is a currently accepted method to determine whether a machine intelligence, the core idea is through verbal communication to determine whether the machine has intelligence.
Turing test can determine whether a machine has intelligence, but its effectiveness is a philosophical question: If a smart AI really have, so he could pretend not to have the intelligence to deceive humans. Turing test have this bug because it is a test of behaviorism, we can not define intelligence from the content, we can only do that (so that the Turing test is a pragmatic approach). Behaviorism and structuralism, we do not understand the mysteries of the human brain generates consciousness and why, so much the difficulty of structuralism, the results of the current model of structuralism have neural network.
How to make artificial intelligence?
Manufacture of artificial intelligence is divided into two schools: Connectionist and behaviorism
Connectionism
Connectionist by imitating something that is manufactured to have an identical function in the field of artificial intelligence, its main contribution is the artificial neural network to achieve intelligent by simulating the human brain, but the current neuroscience do not understand why the human brain will produce consciousness, why is there a smart, we do not fully understand the brain itself, so I want to imitate is not possible. Even if we know the structure of the brain, can imitate, but the cost may be very large and the effect is not necessarily good. Bionics development of mankind tells us that to achieve a certain kind of features does not necessarily exactly the same structure, the same principle as long as you can follow, such as flying, a dream of mankind, we start with the most mechanical structure mockingbird physical structure to manufacture aircraft, but later found not work, but after we have mastered the aerodynamics, but we can create and structure but very different birds flying ability is far greater than the natural bird aircraft. The idea is behaviorism.
Behaviorism
Intelligent does not have to fully mimic the human brain structure, as long as we understand the nature of intelligence, we can create a more powerful agent.
To understand the nature of intelligence, we must first understand what is thinking? Is the search? We are learning?
1. Thinking That search
Information search / search model (Smart Assistant), a major feature is the intelligent search, search the knowledge base according to the needs of our brains that we do not yet know the mechanism
WATSON mechanism of natural language processing information search +
Search requires a good knowledge base, network resource mix, the equivalent of "junk", the current search engines have been able to do a good screening information; mapping knowledge + matching technology, you can use intelligent search.
Search: Find the answers in the existing knowledge
Knowledge come from?
Learning: the discovery of new knowledge
Learning: What is knowledge?
Knowledge is information model (pattern), pattern recognition (upstream research in artificial intelligence), what is the pattern: a stable relationship between information or objects. Such as the three sides of the triangle relationship, the laws of physics (the relationship between the different physical variable) syntax (relationship between different words) languages
When you are making decisions, you are a model of their brain for processing, brain mode (worldview) reflects the objective world, that is your level of knowledge, that determines your decision-making.
Learning is a process of inductive mode (pattern recognition Why so important?)
2. Thinking that is learning
What is the model? Model is speculation mode. Not the algorithm model, model-based algorithm
The entire scientific studies looking mode 1. Determine your research goals, 2. Collect data (experiment), 3. speculation model (try to fit the experimental results, in most cases have to guess a), 4. According to the data model finalized as
Machine learning is to research automation
The basic framework of machine learning
1. The list of possible functions (model) 2. 3. select the best function of the training data
Supervised learning : the tag data requires a lot of labor, very often there is no relevant knowledge (currently the best model is basically supervised model, China's demographic advantage makes it possible to become a powerful artificial intelligence)
Unsupervised Learning : Models summed up their own category (a category not start, the category itself is a kind of knowledge)
Semi-supervised learning : model to be based on sound data marked unlabeled data
Reinforcement Learning : indirect "flag", to evaluate a lot of steps, such as chess game of winners and losers, training dogs and cats (given feedback after a series of actions), reinforcement learning is an iterative process, supervised learning once to completion, giving small amounts of data forming the model, a model output and feedback, then modify the model continuously reinforced.
Rule-based model : defining a relationship between a person and their output characteristics
Based on statistical models who defined features, characteristics and determine the relationship between the output from the model (very important feature:: irrigation papers, pick features to see results; Mongolia industry characteristics is very important, "features works")
Depth learning models : end-to-model data to the data; without defining feature, which can automatically discover features (our own cognitive defects may be), to determine the relationship between the original model and data output
Why deep learning so much?
May find themselves deep learning feature, which uses a multi-layer artificial neural network, the depth at the deep level, breadth and depth that better? The deeper the more complex features can be found (in spite of its high price calculation, there is a risk fitting)
Some important issues of artificial intelligence
Deep learning so much, we can once and for all it? Not work
Artificial mark (labor-intensive), artificial selected characteristics (difficult), choose the model (it is difficult, partly depends Mongolia)
What is a good model? Generalization capability & performance
Good model is between ability and performance find a balance again generalization
Tai Chi Yin and Yang, engineering drawings
Tai Chi Yin and Yang: descriptive very strong, strong generalization ability, but accuracy and computability very poor due to the fitted model
Engineering drawings: poor generalization ability, describe poor calculate specific over-fitting model
Another example is the university philosophy courses provide a less fitting model, meticulous professional courses are over-fitting model, we want to build a better model for understanding the world (worldview), have a lot of self-knowledge
The higher the complexity of the model, the more likely over-fitting (ppt picture is very important), too simple or too complex can not, machine learning technique of looking for the best reflected in VC dimension, making the model generalization and achieve a balance between performance
In order to avoid over-fitting, there is a set of development of the concept
Data sets: the training set, the test set (using the number, the better), the development of sets (number of times is not limited to use)
Using the model developed in the training set to make assessments, unlimited use, but before the model put into application, use an assessment test set to evaluate the results of the test set of high probability model to reflect the true level, because of the test set the probability of frequency of use have been restricted, the model appeared to fit on the test set is small.
Another approach is to maintain a common data set we used for testing, namely common data set
Representative common data set, the quality is relatively high; compare different models and more equitable
(Brush list problems, will still be over-fitting : brush list refers to the common data set will be a good model for other teams draw, the bad model is gradually phased out, and finally the actual multiple teams engaged in a relay race, that is, the best model is ever increasing, ever closer to the common data set, this will still be over-fitting problem)
Over-fitting problem seems simple, but actually very difficult to solve
"Success is the biggest failure" (when this era change , the winners will be more stubborn stick to their inherent experience, but do not want to make a change)
Engineering applications due to cost constraints, often start with a simple model start (such as rule-based model), if not simple model results, try more complex models (such as deep learning model), the following three models in a simple to order complex
Rule-based model: the problem is simple, with a lot of existing knowledge
Based on statistical models: the amount of data, there are some clear characteristics
Depth learning models: no prior knowledge, there is no clear feature, data volume, high power count
Not all problems are suitable for deep learning model, its application of the conditions that must be met; especially large data should not be used more noise when the depth of learning, because it is very sensitive to noise, in this case simple model due to noise sensitivity is low, but the effect is better.
World Affairs
Based on knowledge, knowledge of machine learning. World Knowledge provides accessibility to machine learning
In life, a lot of knowledge can not do without a particular scene or background, such as "Hangzhou, Changchun pharmacy," if there is no background knowledge, it is likely punctuation errors absurd. Machine learning also requires a lot of this background knowledge, this background knowledge to support, assist the learning process, which is referred to as background knowledge.
Small sample learning
People often requires little knowledge can master data (such as human children to learn the language), to achieve a small sample learning actually borrowed external knowledge (knowledge of the world), many small sample of human learning can explain this, except through the children's learning language (current guess is that the human brain has stored knowledge itself, with all the natural language of a high level of abstraction, this knowledge is stored in our genes)
To be continued >>>>